光谱质量、 样本个体差异、 检测系统和建模算法等多种因素共同决定水果糖度检测模型的预测精度和稳定性。 采用自主研发的短积分全透射近红外在线检测系统以5 ms积分时间和0.5 m·s^-1运行速度在线获取了“富士”苹果全透射光谱信号。 不同姿态获取的透射光谱强度差异明显, 但曲线走势相近, 均在920 nm波段具有最大的光谱强度, 在850 nm波段存在波谷。 采用移动平均平滑、 标准正态变量变换和多元散射校正等预处理方法有效去除原始光谱的随机噪声和基线偏差, 减小了样本检测姿态引起的光谱差异。 为分析不同检测姿态对苹果整果糖度预测模型的影响, 构建了单一姿态局限模型和多姿态通用模型, 结果表明基于全位点平均透射光谱构建的单一姿态局限模型对检测姿态具有很大的局限性, 而多姿态通用模型预测能力较单一检测姿态相当, 但却对不同的检测姿态具有更强的适用能力。 为进一步提高光谱信号质量, 优化模型预测能力, 采用信号强度阈值优选方法实现了苹果整果糖度预测模型优化, 发现移除中央位点获取的透射光谱信号, 有利于提高苹果整果糖度预测模型精度。 多姿态通用信号强度优化模型综合考虑不同姿态获取的光谱信息有效性, 有效提升了通用信号强度优化模型的预测能力和稳定性, 当多姿态通用模型中信号强度阈值为5 000时, 模型预测性能最佳, 其预测参数R_p, RMSEP和RPD分别为0.79, 0.84%和1.58。 表明短积分全透射近红外在线检测系统用于不同姿态苹果糖度预测是可行的, 多姿态通用模型的建立, 扩大了模型在不同姿态的预测稳健性, 短积分光谱采集方式结合信号强度阈值优选方法提升了光谱信号的质量和模型的预测能力。

In Vis/NIR nondestructive detection, the accuracy of the prediction model is affected by many factors such as spectral quality, biological variability, detection system and modeling method. In this study, the multi-point full-transmittance spectra (650~1 000 nm) of “Fuji” apple were acquired at a speed of 0.5 m·s^-1 with an integration time of 0.5 ms using an on-line spectrum measurement system. The spectral intensity changed with the detection orientations significantly, but the spectral curves of different orientations were similar, with an obvious peak at 920 nm and an obvious valley at 850 nm. To establish a reliable, accurate, and stable sugar calibration model of intact apple, three spectra preprocessing methods, including moving average smoothing, standard normal variate, and multiplicative scatter correction (MSC), were used to reduce the influence of noise from the environment and instrumental fluctuations. In order to analyze the effect of detection orientations on prediction accuracy, a local model based on single orientation and a universal model based on global orientations were built respectively. The result showed that the prediction accuracy was limited by the detection orientation in the local model, while the universal model had better applicability for multiple detection orientation than that of the local model. In order to further improve the prediction ability, a modeling method named efficient spectrum optimization was proposed to select the spectra with a high signal-to-noise ratio and remove the inefficient transmittance spectrum by investigating the interference of transmittance spectral intensity on the accuracy of the prediction model. The result showed that it is beneficial to optimize the prediction model after removing the spectrum collected from the central zone of the apple. The universal intensity optimization model considered the spectral quality of different orientations comprehensively. The prediction model was best with R_p,RMSEP and RPD of 0.79,0.84% and 1.58 respectively, when the spectral intensity threshold was 5 000. Our result illustrated that the multi-potion spectrum measurement system is promising for on-line detection of apple quality. The modeling method of efficient spectrum optimization could be selective the transmittance spectrum with a high signal-to-noise ratio and optimize the prediction model.